With the widespread application of radar in space target detection and identification, acquiring radar images of non-cooperative targets at different attitudes has become the data foundation for supporting deep learning-based automatic target recognition algorithms. However, the uncontrollable motion trajectories of non-cooperative targets and limited illumination time render the acquisition of measured images extremely difficult; although anechoic chamber measurements and electromagnetic computations can obtain multi-attitude images, they are respectively constrained by high costs and computational time, making it difficult to meet the demand for large-scale data. Existing image generation methods based on Generative Adversarial Networks (GANs) can augment data scales, but most can only control the azimuth angle without precisely constraining the elevation angle, resulting in generated images whose projection morphology does not match reality. Therefore, how to achieve radar image generation with precise dual-dimensional control over both azimuth and elevation angles has become a key technical challenge in data augmentation for radar recognition of non-cooperative targets.
In a recent study published in Space: Science & Technology , a research team from the School of Automation, Central South University, proposed an attitude-controllable radar image generation method for non-cooperative targets based on ControlNet. The study takes the E-3 AWACS aircraft as the target, obtains echo data at various azimuth and elevation angles through electromagnetic computation, and performs imaging; a frequency-domain filtering method is applied to suppress stripe noise, and Lee filtering is adopted to mitigate speckle noise, reducing the image entropy from 3.24 to 3.06. Based on the optical imaging model, the relationship between projection length and elevation angle is derived, and Canny edge maps are used as control conditions input to ControlNet; by freezing the backbone network and training the branch network, attitude-controllable radar image generation is achieved. Experimental results demonstrate that ControlNet-generated images outperform those from InfoGAN and Self-Attention GAN in terms of Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Fréchet Inception Distance (FID), and can still generate clearly contoured images under limited training samples; validation on both E-3 and Su-27 targets further confirms the generalization capability of the proposed method. This research provides an attitude-controllable generation method for augmenting radar image datasets of non-cooperative targets and filling missing aspect angles, offering data support for deep learning-based radar target recognition algorithms.
First, this paper focuses on the data foundation and preprocessing methods for attitude-controllable generation of radar images of non-cooperative targets. The difficulty in acquiring radar images is the primary challenge in non-cooperative target recognition; although electromagnetic computation can generate multi-attitude images, it is time-consuming and suffers from missing aspect angles. To address this, the study establishes a CAD geometric model of the E-3 AWACS aircraft and obtains echo data at various azimuth and elevation angles using the large-element physical optics method at a center frequency of 80 GHz and a bandwidth of 640 MHz. The data are then imaged via 2D inverse Fourier transform, and Rayleigh clutter is superimposed to simulate realistic environments. As illustrated in Fig. 1, the dataset construction workflow comprises four steps: CAD modeling, electromagnetic computation, imaging and preprocessing, and generation of edge maps as control inputs. The electromagnetic imaging results are limited by signal bandwidth and illumination angles; the energy diffusion of strong scattering points forms cross-shaped stripe noise, which severely degrades image quality. The study proposes a frequency-domain filtering method: by analyzing the 2D spectrum, the vertical components corresponding to the stripe noise are located and set to zero for elimination, while Lee filtering is concurrently applied to suppress speckle noise. Fig. 2 presents a comparison of the images before and after denoising; the stripe and speckle noise in the original images are effectively suppressed after processing, with the image entropy reduced from 3.24 to 3.06, thereby validating the effectiveness of the preprocessing. Based on the optical imaging model, the relationship between projection length and elevation angle is derived, and Canny edge maps of optical images at different attitudes are extracted using the Canny operator, providing attitude control conditions for ControlNet.
Second, this paper elaborates on the working principle of ControlNet for attitude-controllable radar image generation. ControlNet employs Stable Diffusion as its backbone network, whose core is a diffusion model that gradually transforms an image into Gaussian noise through forward diffusion, and then recovers the image from random noise through reverse denoising; the training objective is to ensure that the noise predicted by the network matches the noise actually added. However, images generated by the original diffusion model lack fine-grained control capability. As shown in Fig. 3, ControlNet addresses this issue by constructing a branch network: the backbone loads and freezes the pre-trained Stable Diffusion weights to retain its general image generation capability; the branch network, which shares an identical architecture with the backbone's encoder and bottleneck layers, takes edge maps as input and injects the extracted attitude information into the upsampling process of the backbone decoder via zero-convolution modules. The zero convolution is a 1×1 convolution with both initial weights and biases set to zero, ensuring that the branch network does not affect the backbone's generation capability at the early stage of training, while gradually introducing attitude control information as training proceeds. Fig. 4 presents the detailed architecture of the branch network: the 2D convolution module increases the channel dimension; three down-sampling modules extract multi-scale features, with Transformers embedded in the first three groups to enhance global modeling capability; and the bottleneck outputs high-level semantic features. This design enables ControlNet to retain the generation priors learned by the pre-trained model on large-scale optical images, while achieving precise attitude control of radar images through edge maps.
Finally, this paper validates the effectiveness of the proposed method through a series of experiments. Fig. 5 compares the generation results of ControlNet, InfoGAN, and Self-Attention GAN under identical attitudes; the images generated by ControlNet exhibit more accurate detail reproduction in strong scattering point regions (engine, nose, radome), and correctly respond to the dimensional variations induced by changes in elevation angle. On the full dataset of 2,821 images, ControlNet outperforms the comparison methods overall in terms of Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Fréchet Inception Distance (FID). Under limited training samples (only 46 training images), Fig. 6 demonstrates that as training data decrease, the generation quality of InfoGAN continuously degrades, whereas ControlNet shows improvement instead—attributable to the superior distribution modeling capability of diffusion models under low data density and the advantage of pre-training priors. In target self-occlusion scenarios, ControlNet can correctly generate the wing and tail regions occluded by the fuselage, preserving the occlusion relationships consistent with real images. Cross-target generalization experiments on the Su-27 fighter aircraft show that ControlNet-generated images achieve an SSIM of 0.87 and a PSNR of 24.15 dB, surpassing InfoGAN (0.31, 20.82 dB) and SAGAN (0.11, 18.95 dB), thereby verifying the generalization capability of the model. This research provides an attitude-controllable generation method for augmenting radar image datasets of non-cooperative targets and filling missing aspect angles, offering data support for deep learning-based radar target recognition algorithms.
Research on an Attitude Controllable Radar Image Generation Method for Noncooperative Targets
29-Jun-2026